Skip to main content
Connections is where you put credentials that belong to you: a model provider key, a ChatGPT subscription, or environment secrets for the machine.
The Connections screen with no connections yet

A new account has none of these — the platform's own providers are already there

This screen is about models and secrets, not about Gmail or Slack. Connecting an external service — an MCP server, with OAuth or with an API key — happens on the Tools screen.

​
Adding a provider

Add connection offers five:
The provider grid: OpenAI, Anthropic, Gemini, Splox and Codex

What you can connect

The key form is short: a name for the connection, the key itself, and under Advanced a custom base URL when the endpoint is not the provider’s own.
The OpenAI connection form with name, API key and custom base URL

An API-key connection, with the base URL override open

​
What OAuth looks like here

Codex (ChatGPT) is the one that does not take a key. It routes requests through the ChatGPT subscription you already pay for: the device flow opens on chat.openai.com, you authorize there, and Splox keeps the rotating tokens for you afterward. The button says Connect ChatGPT account and there is nothing else to fill in.

​
Where the credential lives

Splox keeps the credential and uses it when a run names the endpoint. What you get in exchange is a new endpoint — the same kind of thing the platform’s own OpenAI, Anthropic, Gemini and Splox entries are. A connected endpoint turns up in two places:
  • In a chat, under the composer’s Advanced → Endpoint, by the name you gave it. Picking it there runs that conversation on your credential.
  • In a harness, as text_llm_endpoint_id on an agent — the id of the endpoint row, and the alternative to provider, never its companion. The ids come from /v2/llm-endpoints.
Each card on the Connections screen shows the provider, the base URL if you set one, and whether it is active. The bin icon disconnects it, after a confirmation; the credential goes with it, and anything pointed at that endpoint has to be pointed somewhere else.

​
Secrets

The second tab is for the machine rather than the model.
The Secrets tab with the environment secrets description

Environment secrets: key/value pairs the sandbox gets as environment variables

Encrypted key/value pairs, injected as environment variables into your agent sandboxes — for the agent and its sub-agents, across every chat. Values are never shown again after you save them. This is where a token belongs when the agent’s own code needs it: a GitHub PAT a script pushes with, an API key a tool of your harness reads from os.environ. Provider keys go in the Providers tab instead, because those are used by the platform, not by your code.

Models

Which model to reach for, and what an agent’s model and provider mean.

Tools

Connecting an external service through an MCP server.